Reuse normalized conversion-report inputs

[chubes4]

Summary

Reporting-maintainability slice under #242: reuse normalized inputs inside ConversionReportProjection::fromResultParts().

• Project fallback diagnostics once instead of three times.
• Normalize/deduplicate runtime islands once instead of three times.
• Build runtime-island summary entries once instead of twice.
• Feed those local arrays to selector/classification summaries and the final report. Preserve the separate summary-row deduplication.

No new class, cache, persistent state, option or public signature. Report keys, ordering, omissions, classifications and deduplication behavior remain unchanged. No unmeasured speed claim.

Verification

Baseline: a57b5f2dd09eaa4a9d81d622a9b4672677516c56.
Candidate: 2d345985af6126ccd4d7a47edf84b3932511a247.

All ten final CI checks passed: PHP 8.2-8.5 package and WordPress integration suites, plus generic visual parity tools, and solved-site promotion.

• New behavioral test checks overlapping runtime declarations, first-occurrence ordering, selector/source order, fallback order, and exact classification/preservation counts.
• Full PHP 8.4 / 512 MB Composer gate passed: canonical, unit, 305 parity fixtures, dist shape and install proof.
• Coordinator independently reran the full 385-file HTML corpus and nine targeted HTML/artifact cases. Complete envelope/block comparison records match byte-for-byte, excluding only recursive duration; final record SHA-256 520a6c71343bab56933991a5cd51cc85483207c03f0204d37998b8e35f1bc6f4.
• The full/compact invariant tool passed on 385 fixtures, including 365 warning documents, with identical operational outputs and only declared detail differences.
• WordPress integration and solved-site promotion passed separately in final CI as linked above; they are not claimed from standalone lab tests.

Reproduce

From php-transformer, with dependencies and child PHP memory configured:

composer test
php tests/unit/conversion-report-projection-input-reuse.php
php -d memory_limit=512M tools/benchmarks/validation-evidence-compare.php ../fixtures
REQUIRE_WP_TESTS=1 WP_TESTS_DIR=/path/to/wordpress-tests composer test:wordpress-integration

The corpus comparison command from #1627 applies unchanged to the baseline/candidate revisions above with matching dependencies.

AI Assistance

OpenAI GPT-5.6 Terra via direct OpenCode implemented the local reuse and behavioral test, then ran lab gates. GPT-6 Astra via OpenCode coordinated and reviewed ordering/count/deduplication semantics, independently reran the final corpus/artifact comparison and prepared this PR under Chris Huber's direction. No release or deployment was performed.